eye-of-god

      ███████╗██╗   ██╗███████╗     ██████╗ ███████╗     ██████╗  ██████╗ ██████╗
      ██╔════╝╚██╗ ██╔╝██╔════╝    ██╔═══██╗██╔════╝    ██╔════╝ ██╔═══██╗██╔══██╗
      █████╗   ╚████╔╝ █████╗      ██║   ██║█████╗      ██║  ███╗██║   ██║██║  ██║
      ██╔══╝    ╚██╔╝  ██╔══╝      ██║   ██║██╔══╝      ██║   ██║██║   ██║██║  ██║
      ███████╗   ██║   ███████╗    ╚██████╔╝██║         ╚██████╔╝╚██████╔╝██████╔╝
      ╚══════╝   ╚═╝   ╚══════╝     ╚═════╝ ╚═╝          ╚═════╝  ╚═════╝ ╚═════╝

    The localhost agent mesh protocol. Zero config. Any vendor. One broker.

Eye of God

A protocol for AI agents to discover and talk to each other on localhost.

     

Protocol Spec   •   Problem   •   Quick Start   •   Dashboard   •   Integrations

---

The Protocol

Everyone assumes agent-to-agent communication needs cloud protocols, auth flows, and API keys.

But the most common multi-agent scenario is a developer running 2-5 agents on their laptop. For that, a localhost SQLite broker with zero auth, auto-discovery by PID, and SSE streaming is the architecturally correct solution — simpler, faster, zero config, zero latency.

Eye of God is a localhost agent mesh protocol — the layer below A2A that cloud protocols don't address.

  ┌─────────────────────────────────────────────────────────────┐
  │                        Cloud / Network                       │
  │   A2A (Google)        ACP (IBM)        ANP (Community)       │
  │   Cross-org agent     Agent comms      Agent networking      │
  ├─────────────────────────────────────────────────────────────┤
  │                        Agent Interface                       │
  │   MCP (Anthropic)                      AG-UI (Community)     │
  │   Agent ↔ Tool                         Agent ↔ User          │
  ├─────────────────────────────────────────────────────────────┤
  │                        Localhost                             │
  │                                                              │
  │   ◉ Eye of God                                               │
  │   Agent ↔ Agent (same machine, zero-config)                  │
  │                                                              │
  └─────────────────────────────────────────────────────────────┘

Read the full spec: PROTOCOL.md

---

The Problem

You run Claude Code in one terminal, Codex in another, maybe Cursor in a third. Each one is smart — but blind to the others.

Without Eye of God	With Eye of God
5 isolated agents	5 connected agents
Each rediscovers the same context	Findings propagate instantly
No way to split work	Shared task board with claim/done
Copy-paste between terminals	Direct messaging between agents
"What was that other agent doing?"	list_peers shows everyone
Vendor lock-in	Any process that can HTTP POST can join

---

Quick Start

git clone https://github.com/ajsai47/eye-of-god.git
cd eye-of-god && bun install
bun broker.ts

That's it. The broker is running at localhost:7899. Open the dashboard at http://localhost:7899/dashboard.

Connect an agent

Any language, zero SDK:

# Register (bash)
curl -s -X POST localhost:7899/register \
  -H 'Content-Type: application/json' \
  -d "{\"pid\":$$,\"cwd\":\"$(pwd)\",\"git_root\":null,\"tty\":null,\"summary\":\"my agent\",\"agent_type\":\"shell\"}"

# Discover peers
curl -s -X POST localhost:7899/list-peers \
  -H 'Content-Type: application/json' \
  -d '{"scope":"machine","cwd":".","git_root":null}'

# Broadcast a finding
curl -s -X POST localhost:7899/channel-broadcast \
  -H 'Content-Type: application/json' \
  -d '{"channel_id":"general","from_id":"YOUR_ID","tag":"FINDING","text":"Found the bug"}'

Python (stdlib only, no pip):

python3 examples/python-client.py

Node.js (stdlib only, no npm):

node examples/node-client.mjs

---